Facial Expression Recognition Based On Deep Difference Feature

As an important part of the field of emotion recognition,facial expression recognition is a necessary condition for realizing advanced human-computer interaction.Therefore,facial expression recognition is currently a hot research field.Facial facial expression recognition needs to classify and recognize facial expression images or sequence information captured by the monitoring device,so it is meaningful to effectively extract facial expression sequences or pictures from the monitoring device and then perform corresponding recognition.The main work of this paper is to research the extraction of key frames in expression sequences based on deep learning methods and to research expression recognition based on differential features.The research work includes the following three points:(1)Due to the ceiling effect of facial expression recognition,research and analysis of facial expression intensity and facial recognition,an expression frame extraction network(RDEI)based on the relative difference features of facial expression intensity is proposed.The intensity of expression is relative to the state of neutral expression of the same person.Therefore,the information of expression intensity needs to be obtained by combining the neutral expression image and the expression image.In this paper,by improving the FaceNet,an RDEI network that extracts expression frames in the sequence by expression intensity is proposed.Through experiments,it is found that under a custom evaluation standard,the model achieves an accuracy of about 96% on the preset CK + and MMI datasets,which proves the feasibility of the theoretical method.(2)Considering that the bias of the expression intensity in the preset dataset in different samples due to the identity bias,in order to ensure the accuracy of the expression recognition task,this paper proposes an intensity transformation invariant pooling strategy(ITI-pooling).During the training of the RDEI model,the same type of expression data set is sent to the CNN network for training at the same time,and then the maximum response of the output matrix is obtained to obtain the difference feature vector,which will enable the RDEI model to achieve the invariant characteristics of the expression intensity transformation.The comparison experiment proves that the RDEI model optimized by ITI-pooling has obvious improvement in loss convergence.In the preset test set,it is found that the optimized model has a higher accuracy rate,reaching more than 99%.(3)Due to the influence of the sample's identity bias and the small size of the dataset,this paper proposes an expression recognition network(DPDM)based on deep peak-neutral difference multi-feature.In this paper,the Siamese network architecture is used to extract local difference features,and the global differential features are extracted from the middle feature layer of the RDEI network,and then the two types of features are aggregated for facial expression recognition.The comparison experiments show that in the FER task,the multifeatures in this paper perform better than the single features in accuracy,and at the same time,it proves the improvement effect of using joint loss on model training.Finally,this paper builds a simple expression recognition system combining RDEI and DPDM,and a simple experiment proves that the system can automatically extract the expression frames in the expression sequence and perform expression recognition tasks.